Narrative reasoning relies on the understanding of events in story contexts, which requires abundant background world knowledge. To help machines leverage such knowledge, existing solutions fall into two categories. Some focus on implicitly modeling event knowledge by pretraining language models (LMs) with event-aware objectives, which breaks down knowledge structures and suffers from poor interpretability. Others explicitly collect world knowledge of events into structured event-centric knowledge graphs (KGs). However, existing research on leveraging these knowledge sources for free-texts is largely absent. In this thesis, we propose an initial comprehensive framework, Even-Ground, to tackle the problem of grounding¹ free-texts to event-centric KGs for contextualized narrative reasonin...[ Read more ]

Narrative reasoning relies on the understanding of events in story contexts, which requires abundant background world knowledge. To help machines leverage such knowledge, existing solutions fall into two categories. Some focus on implicitly modeling event knowledge by pretraining language models (LMs) with event-aware objectives, which breaks down knowledge structures and suffers from poor interpretability. Others explicitly collect world knowledge of events into structured event-centric knowledge graphs (KGs). However, existing research on leveraging these knowledge sources for free-texts is largely absent. In this thesis, we propose an initial comprehensive framework, Even-Ground, to tackle the problem of grounding¹ free-texts to event-centric KGs for contextualized narrative reasoning. We point out two critical problems on this direction, namely the event representation and sparsity problems, and address them by simple while effective parsing and abstraction methods. Experimental results on several representative narrative reasoning tasks that require commonsense reasoning ability show that our approach consistently outperforms baseline models and achieves new state-of-the-art performance, while providing human-interpretable evidence.

¹Here, the term “grounding” refers to a process similar to “linking” used in “entity linking”, where the target is the event-centric KGs.